Leveraging Electronic Medical Records for Early Lung Cancer Diagnosis.

introduction

Background

In the US, only 27.4% of lung cancer cases are diagnosed early, with 5-year survival rates of 63% for localized and 27% for late-stage cancers (1). Despite recommendations since 2012 for screening high-risk individuals with low-dose CT, uptake has been limited, and most lung cancer diagnoses occur after symptoms appear (2). Similarly, chest x-rays, while commonly used as an initial test to investigate patients with symptomatic suspected lung cancer, have demonstrated limited sensitivities between 50-70% and specificities over 80% (3).

Symptoms often overlap with common conditions, making detection challenging, and studies have identified median delays of 187 days from symptom onset to diagnosis. This prolonged interval presents an opportunity for improvement.

This study examines the use of the AI cancer prediction platform, C the Signs, to passively screen for lung cancer by leveraging electronic medical records (EMRs) for early lung cancer detection.

METHODS

Study Design

This retrospective study aimed to evaluate the performance of the C the Signs AI Cancer Prediction Model in identifying individuals at high risk of lung cancer earlier than standard physician-led diagnostic pathways. Using the Mayo Data Platform, we accessed electronic medical records (EMRs) for a cohort of 894,409 patients, spanning a 20-year period from January 1, 2002 to December 31, 2021. A total of 7,395 patients were diagnosed with lung cancer during this timeframe.

The study included adults aged ≥18 years with at least one recorded primary care interaction. Patients with a diagnosis of lung cancer prior to the study period were excluded. The AI model was retrospectively applied to the EMR dataset to identify individuals at risk of lung cancer and assess its ability to flag these cases earlier than traditional clinical pathways.

{{feedback}}

Model Application

The C the Signs AI Prediction Model is a pan-cancer expert system developed in the United Kingdom to assist healthcare professionals in identifying patients at risk of cancer, including lung cancer, at the point of care. The model integrates a combination of risk factors, symptoms, and clinical signs, employing a dual-layered algorithm that combines structured data, such as coded clinical information (e.g., ICD-10 codes), with unstructured data derived from EMRs using Natural Language Processing (NLP) techniques. This innovative approach enhances the model’s ability to identify subtle risk indicators that might be overlooked in traditional diagnostic pathways.

{{feedback-2}}

Analysis

We assessed the model’s diagnostic performance using standard metrics: sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). We also compared the time of AI-predicted risk identification to the date of clinical diagnosis made by primary care physicians to evaluate whether the AI platform facilitated earlier detection. Although lung cancer staging data was not available in this dataset, the timing comparison provided a proxy for assessing the model’s potential impact on accelerating diagnosis and pathway entry.

results

Model Performance

The C the Signs AI cancer prediction platform demonstrated a sensitivity of 91.5% and a specificity of 52.3% in identifying patients at risk of lung cancer. Among the 7,395 patients diagnosed with lung cancer in the dataset, the model correctly flagged 6,749 cases.
Notably, the platform identified 26.6% of patients at risk up to five years earlier than when their diagnoses were made by primary care physicians, underscoring its potential to support earlier detection in lung cancer pathways.

Of the 894,409 patients eligible in the dataset, 423,249 false positives were recorded among the 887,014 patients without lung cancer, contributing to the specificity result.

This study analyzes de-identified Electronic Health Record (EHR) data via Mayo Clinic Platform Discover. The data presented here is extracted following a privacy-preserving protocol
and deemed de-identified by an expert in accordance with HIPAA Privacy Rule. Additional data, including raw EHR data, cannot be shared due to expert determination. Contact corresponding authors for more information on Mayo Clinic Platform Discover.

Model Performance

The C the Signs AI cancer prediction platform demonstrated a sensitivity of 91.5% and a specificity of 52.3% in identifying patients at risk of lung cancer. Among the 7,395 patients diagnosed with lung cancer in the dataset, the model correctly flagged 6,749 cases.
Notably, the platform identified 26.6% of patients at risk up to five years earlier than when their diagnoses were made by primary care physicians, underscoring its potential to support earlier detection in lung cancer pathways.

Of the 894,409 patients eligible in the dataset, 423,249 false positives were recorded among the 887,014 patients without lung cancer, contributing to the specificity result.

Model Performance

The C the Signs AI cancer prediction platform demonstrated a sensitivity of 91.5% and a specificity of 52.3% in identifying patients at risk of lung cancer. Among the 7,395 patients diagnosed with lung cancer in the dataset, the model correctly flagged 6,749 cases.
Notably, the platform identified 26.6% of patients at risk up to five years earlier than when their diagnoses were made by primary care physicians, underscoring its potential to support earlier detection in lung cancer pathways.

Of the 894,409 patients eligible in the dataset, 423,249 false positives were recorded among the 887,014 patients without lung cancer, contributing to the specificity result.

This study analyzes de-identified Electronic Health Record (EHR) data via Mayo Clinic Platform Discover. The data presented here is extracted following a privacy-preserving protocol
and deemed de-identified by an expert in accordance with HIPAA Privacy Rule. Additional data, including raw EHR data, cannot be shared due to expert determination. Contact corresponding authors for more information on Mayo Clinic Platform Discover.

Model Performance

The C the Signs AI cancer prediction platform demonstrated a sensitivity of 91.5% and a specificity of 52.3% in identifying patients at risk of lung cancer. Among the 7,395 patients diagnosed with lung cancer in the dataset, the model correctly flagged 6,749 cases.
Notably, the platform identified 26.6% of patients at risk up to five years earlier than when their diagnoses were made by primary care physicians, underscoring its potential to support earlier detection in lung cancer pathways.

Of the 894,409 patients eligible in the dataset, 423,249 false positives were recorded among the 887,014 patients without lung cancer, contributing to the specificity result.

This study analyzes de-identified Electronic Health Record (EHR) data via Mayo Clinic
Platform Discover. The data presented here is extracted following a privacy-preserving protocol
and deemed de-identified by an expert in accordance with HIPAA Privacy Rule. Additional data, including raw EHR data, cannot be shared due to expert determination. Contact corresponding authors for more information on Mayo Clinic Platform Discover.

• The C the Signs Al cancer prediction platform demonstrated a sensitivity of 91.5% and a specificity of 52.3% in identifying patients at risk of lung cancer.
• Among the 7,395 patients diagnosed with lung cancer in the dataset, the model correctly flagged 6,749 cases.
• Notably, the platform identified 26.6% of patients at risk up to five years earlier than when their diagnoses were made by primary care physicians, underscoring its potential to support earlier detection in lung cancer pathways.
• Of the 894,409 patients eligible in the dataset, 423,249 false positives were recorded among the 887,014 patients without lung cancer, contributing to the specificity result.

interpreting the results

Discussion

This study demonstrates the potential of the C the Signs AI Prediction Model in improving early detection of lung cancer. With a sensitivity of 91.5%, the model performs comparably to or better than current clinical tools, such as chest X-ray, which are often used reactively and have limited sensitivity for early-stage disease.

The AI platform’s performance is notable in its ability to detect 26.6% of lung cancer cases up to five years earlier than the diagnoses made through traditional physician-led pathways, supporting its value in identifying patients before significant clinical progression occurs.

While the specificity of 52.3% suggests a moderate false-positive rate, this aligns with trade-offs seen in other early detection models prioritising high sensitivity. Importantly, the platform’s high NPV indicates strong potential to safely rule out low-risk patients, potentially reducing unnecessary imaging, referrals, or anxiety.

One of the key advantages of this AI model is its capacity to flag individuals who might not yet meet established clinical referral thresholds or who present with non-specific symptoms. This capability is particularly relevant in lung cancer, where early symptoms can be vague or absent, and diagnosis often occurs at an advanced stage.

By identifying risk earlier, the model supports more timely investigation and intervention,
with implications for improved survival and reduced morbidity. These benefits may outweigh the cost of additional diagnostics prompted by early risk identification.